LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving

TL;DR

LaST-VLA introduces a physically grounded latent reasoning framework for autonomous driving, improving spatial-temporal understanding and safety by integrating geometric constraints and dynamic foresight into a unified model.

Contribution

It proposes a novel Latent Spatio-Temporal Chain-of-Thought approach that incorporates geometric and dynamic constraints into latent reasoning for autonomous driving.

Findings

Achieved new state-of-the-art on NAVSIM v1 and v2 benchmarks.

Excelled in spatial-temporal reasoning on SURDS and NuDynamics.

Enhanced safety and rule compliance through reinforcement learning.

Abstract

While Vision-Language-Action (VLA) models have revolutionized autonomous driving by unifying perception and planning, their reliance on explicit textual Chain-of-Thought (CoT) leads to semantic-perceptual decoupling and perceptual-symbolic conflicts. Recent shifts toward latent reasoning attempt to bypass these bottlenecks by thinking in continuous hidden space. However, without explicit intermediate constraints, standard latent CoT often operates as a physics-agnostic representation. To address this, we propose the Latent Spatio-Temporal VLA (LaST-VLA), a framework shifting the reasoning paradigm from discrete symbolic processing into a physically grounded Latent Spatio-Temporal CoT. By implementing a dual-feature alignment mechanism, we distill geometric constraints from 3D foundation models and dynamic foresight from world models directly into the latent space. Coupled with a progressive SFT training strategy that transitions from feature alignment to trajectory generation, and refined via Reinforcement Learning with Group Relative Policy Optimization (GRPO) to ensure safety and rule compliance. \method~setting a new record on NAVSIM v1 (91.3 PDMS) and NAVSIM v2 (87.1 EPDMS), while excelling in spatial-temporal reasoning on SURDS and NuDynamics benchmarks.